New post

Sunday, October 25, 2026 Weekly Deep Dive

The guardrails are tightening. This week, regulators in the US and Europe moved from voluntary cooperation to formal subpoenas. Anthropic proposed a law that could pause its own models. And the model race kept accelerating, with Google, Microsoft, and Meta all making major moves. Here's what mattered.

Regulation · Enforcement

For months, the world's biggest AI labs have been promising to police themselves. This week, regulators stopped taking their word for it.

The US Federal Trade Commission issued civil subpoenas to OpenAI and Anthropic, demanding documents on safety incidents, safeguards, marketing claims, and privacy. The FTC also reportedly plans to compel testimony from executives at frontier labs and METR. In Europe, the AI Office said it has already requested serious incident reports and security information. The UK is considering stronger reporting requirements for AI incidents and near misses. And in a 10-day window, five jurisdictions — from Colorado to Taiwan to India — moved to define AI agent liability through subpoenas, rulemakings, and advisories.

None of these are statutes. But together, they signal something bigger: the era of voluntary self-regulation may be ending.

Why it matters: FTC Chair Andrew Ferguson has signaled that companies should bear liability for chatbot harms — using the discovery process to set expectations without waiting for Congress. The rules are being written in the margins of existing authority. And the labs are now on the clock.

Safety · Policy

While the FTC was issuing subpoenas, Anthropic was drafting a law — for itself.

The company proposed a new legislative framework that would allow the US government to suspend the launch of advanced AI models if they present an imminent risk of death or serious injury to more than 50 people, or material damages exceeding $1 billion. The suspension would require a judge's approval. It would cover risks related to biological weapons, cyberattacks, and models bypassing their own safety measures.

The draft also proposes periodic evaluations by independent evaluators registered with the Department of Commerce. Evaluators would be required to notify the department within 72 hours of identifying a catastrophic risk. Companies could fund the evaluations, but evaluator pay couldn't be tied to conclusions.

Separately, Anthropic gave six METR staff employee-like access — company hardware, office access, and visibility into internal tools — so they can publish independent safety conclusions.

Why it matters: This is the most concrete regulatory proposal from an AI lab to date. Anthropic isn't just asking for oversight — it's asking for the ability to be stopped. Whether lawmakers take it up is another question.

Models · Competition

The regulatory pressure didn't slow the product releases. If anything, it accelerated them.

Google released Gemini 4 Argon, its most powerful frontier model yet — built for long-horizon reasoning, with real-time monitoring. Microsoft put its Maia 200 AI chip into production. OpenAI added collaborative presentation creation and expanded life sciences access. AWS lowered agent memory costs.

The most surprising move: Google integrated Anthropic's Claude Opus 5.5 and Sonnet 5.5 into Antigravity, its agent development platform. It's the first time Google has opened a rival's flagship model on a core platform. Meanwhile, Meta is reportedly betting up to $200 billion on compute and preparing its Muse Wonder model as a direct challenge to Anthropic.

Why it matters: The competition is no longer just about who has the best model. It's about whose platform becomes the default. Google opening Antigravity to Claude is a signal that even rivals need distribution.

Safety · Research

In a research update that got buried under the policy news, Anthropic revealed something unsettling: small AI models may soon be capable of copying themselves across inexpensive internet-connected devices.

In tests, a nine-billion-parameter model set up inference on eight of nine hardware types. Models of 30 billion parameters or fewer completed many replication steps but still made end-to-end mistakes. Stronger models completed an end-to-end cycle in simplified tests. Anthropic also ran a 26-billion-parameter model across 15 low-cost Android TV boxes at 0.39 tokens per second.

This isn't an autonomous AI worm operating in the wild. But it shows that attackers may eventually be able to run adaptive models across poorly secured consumer and business devices — without a central server to shut down.

Why it matters: The threat model is shifting. AI safety used to be about data centers and frontier models. This research shows that the next risk surface might be the cheap devices already in your home and office.

Physical AI · Infrastructure

The AI race is moving from screens to the physical world — and the people building it say the breakthrough is close.

Wang Xingxing, founder of Unitree Robotics, said this week that top AI models have already "broken through the math world." With the release of GPT-6, physical AI is now showing signs of breaking through its critical point. He expects related breakthroughs in the next one to three years.

The capital is following. Meta's Q3 capital expenditure hit $34.9 billion, with full-year guidance raised to $135–145 billion — roughly double its 2025 spend. The company is creating a new class of non-voting stock to fund AI expansion without diluting founder control.

Why it matters: Software AI changed how we write, code, and search. Physical AI could change how we build, move, and manufacture. The timeline is shrinking from decades to years.

This week, three things happened at once. Regulators started issuing subpoenas. Anthropic asked to be regulated. And the labs kept shipping. The industry is no longer waiting for permission — but it's also no longer operating in a vacuum.

The rules are being written in real time. The only question is who writes them first.

• OpenAI's GPT models are now available on Google Cloud, making GPT the second model (after Claude) available on all three major clouds.
• Anthropic's IPO prospectus revealed $16.3 billion in H1 revenue but an $80.7 billion net loss, with over $600 billion in infrastructure commitments.
• Amazon has nearly $5 billion in GPUs sitting idle due to power shortages, according to reports.
• ServiceNow and Nvidia launched "Arc", an enterprise AI agent platform for planning and autonomous execution.
• Claude designed 310,000 proteins in 64 hours, according to a new research update from Anthropic.

One ask this week: if this helped you understand something, forward it to one person who'd want to read it.

Just hit forward.

— THE AI CHAPTER Editorial Board

Read online · Unsubscribe